Characterizing and mapping cropping patterns in a complex agro-ecosystem: An iterative participatory mapping procedure using machine learning algorithms and MODIS vegetation indices

Accurate and up-to-date spatial agricultural information is essential for applications including agro-environmental assessment, crop management, and appropriate targeting of agricultural technologies. There is growing research interest in spatial analysis of agricultural ecosystems applying satellite remote sensing technologies. However, usability of information generated from many of remotely sensed data is often constrained by accuracy problems. This is of particular concern in mapping complex agro-ecosystems in countries where small farm holdings are dominated by diverse crop types. This study is a contribution to the ongoing efforts towards overcoming accuracy challenges faced in remote sensing of agricultural ecosystems. We applied time-series analysis of vegetation indices (Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI)) derived from the Moderate Resolution Imaging Spectrometer (MODIS) sensor to detect seasonal patterns of irrigated and rainfed cropping patterns in five townships in the Central Dry Zone of Myanmar, which is an important agricultural region of the country has been poorly mapped with respect to cropping practices. To improve mapping accuracy and map legend completeness, we implemented a combination of (i) an iterative participatory approach to field data collection and classification, (ii) the identification of appropriate size and types of predictor variables (VIs), and (iii) evaluation of the suitability of three Machine Learning algorithms: Support Vector Machine (SVM), Random Forest (RF), and C5.0 algorithms under varying training sample sizes. Through these procedures, we were able to progressively improve accuracy and achieve maximum overall accuracy of 95% When a small sized training dataset was used, accuracy achieved by RF was significantly higher compared to SVM and C5.0 (P < 0.01), but as sample size increased, accuracy differences among the three machine learning algorithms diminished. Accuracy achieved by use of NDVI was consistently better than that of EVI (P < 0.01). The maximum overall accuracy was achieved using RF and 8-days NDVI composites for three years of remote sensing data. In conclusion, our findings highlight the important role of participatory classification, especially in areas where cropping systems are highly diverse and differ over space and time. We also show that the choice of classifiers and size of predictor variables are essential and complementary to the participatory mapping approach in achieving desired accuracy of cropping pattern mapping in areas where other sources of spatial information are scarce.